Content-based image retrieval systems use low-level features likecolor and texture for image representation. Given these representationsas feature vectors, similarity between images is measured by computingdistances in the feature space. Unfortunately, these low-level featurescannot always capture the high-level concept of similarity in humanperception. Relevance feedback tries to improve the performance byallowing iterative retrievals where the feedback information from theuser is incorporated into the database search. We present a weighteddistance approach where the weights are the ratios of standarddeviations of the feature values both for the whole database and alsoamong the images selected as relevant by the user. The feedback is usedfor both independent and incremental updating of the weights and theseweights are used to iteratively refine the effects of different featuresin the database search. Retrieval performance is evaluated using averageprecision and progress that are computed on a database of approximately10,000 images and an average performance improvement of 19% is obtainedafter the first iteration
展开▼
